A Deep Neural Network Model for Realtime Semantic-Segmentation Video Processing supported to Autonomous Vehicles

Traffic congestion has been a huge problem, especially in urban area during peak hours, which causes a major problem for any unmanned/autonomous vehicles and also accumulate environmental pollution. The solutions for managing and monitoring the traffic flow is challenging that not only asks for performing accurately and flexibly on routes but also requires the lowest installation costs. In this paper, we propose a synthetic method that uses deep learning-based video processing to derive density of traffic object over infrastructure which can support usefull information for autonomous vehicles in a smart control system. The idea is using the semantic segmentation, which is the process of linking each pixel in an image to a class label to produce masked map that support collecting class distribution among each frame. Moreover, an aerial dataset named Saigon Aerial with more than 110 samples is also created in this paper to support unique observation in a biggest city in Vietnam, HoChiMinh city. To present our idea, we evaluated different semantic segmentation models on 2 datasets: Saigon Aerial and UAVid. Also to track our model’s performance, F1 and Mean Intersection over Union metrics are also taken into account. The code and dataset are uploaded to Github and Kaggle repository respectively as follow: Saigon Aerial Code, Saigon Aerial dataset